acceptodds
Under review as a conference paper at ICLR 2027

RAFMa: Retrieval-Augmented Flow Matching for Global Visual Geolocation

Abstract

Global visual geolocation predicts where on Earth an image was taken. Classification and generative models store all world knowledge in their parameters, cannot be updated without retraining, and are known to generalize well at coarse scales, such as country or region. In contrast, retrieval-based systems compare the image against a database of geotagged images and excel at fine scales, such as street or building, where memorization matters; moreover, their database can be updated at any time as new geotagged images become available. We combine the two families in a flow matching model conditioned jointly on the query image and on retrieved visually similar images together with their geotags, a form of in-context learning. Starting from a random location on the sphere, the model iteratively moves it along a learned velocity field, while an attention module learns to weight each retrieved image according to its appearance and geodesic distance to the current estimate. The model achieves a good trade-off between generalization and memorization, and effectively benefits from knowledge updates at inference time by changing the database. On street-view images (OSV-5M), RAFMa sets a new state of the art, clearly outperforming prior work. On unconstrained user photographs (YFCC4k, Im2GPS3k), it is competitive with systems built on multimodal LLMs of several billion parameters, while training only 135M parameters.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
visual geolocation, image geolocalization, retrieval-augmented generation, flow matching, generative models, image retrieval

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_W2sbOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces RAFMa, a retrieval-augmented model for global visual geolocation. It extends spherical flow matching by conditioning the generation process on a set of visually similar geotagged images retrieved from an external database. A context module combines visual information from the retrieved images with their geographic coordinates, while a time-dependent geodesic bias adjusts their influence according to the current location hypothesis during generation. The method is evaluated on OSV-5M, YFCC4k, and Im2GPS3k using several frozen image encoders. RAFMa consistently improves over matched context-free flow-matching baselines and achieves particularly strong results on OSV-5M. Additional experiments study the contributions of the geographic tokens and distance bias, the effects of training-set and database size, context quality, and performance under shared retrieval candidates with GeoRanker.

Critical strengths
  1. The proposed integration of retrieval and generative geolocation is technically well motivated. RAFMa does more than select or rerank a retrieved location. Its prediction can lie outside the retrieved candidates, and the relevance of each context image changes with the evolving geographic hypothesis. The geodesic attention bias is a natural inductive bias for this setting and distinguishes the method from simpler retrieval aggregation.
  2. The empirical validation is unusually thorough and generally convincing. The paper includes matched context-free baselines across multiple encoders, component ablations, controlled variation of training and database size, analysis of context quality, and a shared-retrieval comparison with GeoRanker. The experiment excluding all database images within 1 km of the query is also useful because it shows that the gains at broader geographic scales are not explained entirely by retrieving a nearby image.
Critical weaknesses

The main limitation is that the interpretation of the finest-scale results is complicated by the additional snapping post-processing. On the Flickr benchmarks, this step substantially improves 1 km accuracy while having almost no effect at broader scales. In contrast, it reduces 1 km accuracy on OSV-5M. Since the snapping radius is selected on training queries, it is unclear how robust this heuristic is across datasets with different retrieval coverage. This does not undermine the improvements of the underlying RAFMa model, but makes it difficult to attribute the reported fine-scale gains entirely to the proposed generative approach.

Questions for authors

Could the authors explain why the snapping radius is selected on training queries and why post-processing is retained on OSV-5M despite reducing its 1 km accuracy? How sensitive are the finest-scale results to this radius?

2 comments
AuthorsOpenReview ↗

Coordinate snapping

We thank the reviewer for the careful reading and address the concern about the post-processing below.

Weakness. The main limitation is that the interpretation of the finest-scale results is complicated by the additional snapping post-processing. On the Flickr benchmarks, this step substantially improves 1 km accuracy while having almost no effect at broader scales. In contrast, it reduces 1 km accuracy on OSV-5M. Since the snapping radius is selected on training queries, it is unclear how robust this heuristic is across datasets with different retrieval coverage. This does not undermine the improvements of the underlying RAFMa model, but makes it difficult to attribute the reported fine-scale gains entirely to the proposed generative approach.

Question. Could the authors explain why the snapping radius is selected on training queries and why post-processing is retained on OSV-5M despite reducing its 1 km accuracy? How sensitive are the finest-scale results to this radius?

We select the snapping radius of the post-processing on training queries so that no setting depends on the test images. The tuning rule was: take the radius with the largest gain at 1 km among those that lose at most 0.1 points at 25 km, which gives 2 km on OSV-5M and 10 km on MP-16. This rule is described in Appendix A (paragraph "Prediction post-processing"), where we have now added Figure 6, showing the effect of the radius on the training queries and on the test sets.

On the Flickr benchmarks, the training and test curves have the same shape. The gain at 1 km grows up to snapping radius of about 10 km and then saturates, so the results are not sensitive to the radius. At 1 km, the selected radius gives performance within 1.2 points of the best radius per test set.

On OSV-5M, the selection on training queries does not transfer well to the test set; see Table A1 in our answer to Reviewer Xogd. The benchmark keeps every test image at least 1 km away from the database images, so a prediction snapped to one of them is never within 1 km of the true location. The effect is small since the chosen radius is 2 km. Accuracy at 1 km drops from 0.55% to 0.16%, while accuracy at 25 km and GeoScore remain practically unchanged. We kept the post-processing on both datasets to be fair and not to tune the design choices based on test set performance.

Thank you for the detailed clarification and additional sensitivity analysis. The explanation of the snapping radius selection and its behavior on OSV-5M addresses my main concern.

After examining the results without post-processing in Tables 7 and 8, I find that RAFMa's core retrieval-conditioned generation mechanism provides meaningful improvements independently of coordinate snapping, particularly at broader geographic scales. While coordinate snapping contributes substantially to the finest-scale results on the Flickr benchmarks, the results without post-processing also demonstrate the effectiveness of RAFMa's core approach.

Overall, I find the proposed method technically well motivated and the experimental evaluation convincing. I have therefore decided to raise my rating.

Reviewer_TAsKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RAFMa, a visual geolocation model that conditions spherical flow matching on retrieved geotagged images. The model combines visual self-attention and geographic cross-attention, with a geodesic distance bias that adjusts context weights according to the location being generated. Experiments on OSV-5M and two Flickr benchmarks examine localization accuracy, context quality, database size, and the contributions of individual components.

Critical strengths
  1. Combining the generalization of a generative model with the fine-grained evidence and updatable memory provided by retrieval is well motivated and practically useful.

  2. Spherical flow matching is a reasonable choice for predicting continuous locations, and using the evolving location to adjust context weights connects retrieval with the generation process.

  3. The matched-encoder comparisons, ablations, and database expansion experiments support the main claims about accuracy improvements, the value of geographic context, and the ability to use new examples without retraining.

Critical weaknesses
  1. The framework builds on existing spherical flow matching, visual retrieval, and attention mechanisms, with the main additions concentrated in contextual conditioning and the geodesic distance bias. The paper should distinguish these contributions more clearly from prior retrieval-augmented geolocation methods and explain what modeling capabilities they add beyond standard conditioning.

  2. Retrieval relies on frozen pretrained descriptors and cosine similarity, with little task-specific design. There is room to improve the retrieval representation or matching mechanism through local feature correspondences, geometric consistency, and greater robustness to viewpoint and lighting changes, reducing reliance on global appearance similarity.

  3. The number of trainable parameters does not establish the efficiency of the full system. The reported 135M parameters exclude the frozen visual encoder, while inference also requires retrieval from a large feature database and 250 guided sampling steps. Without comparisons of end-to-end inference time, peak GPU memory, and database storage, the practical accuracy–cost trade-off remains unclear.

Questions for authors
  1. What are the main modeling advances over existing spherical flow matching and retrieval-augmented geolocation methods? What does the geodesic distance bias add beyond ordinary conditional attention?

  2. What motivates the choice of frozen descriptors and cosine retrieval? How could local geometric information and invariance to viewpoint and lighting changes be incorporated into a more task-specific retrieval mechanism?

  3. Please report the full pipeline’s inference time, peak GPU memory, and database storage requirements, including visual encoding, retrieval, and generation, and discuss how accuracy and cost change with database size and sampling steps.

4 comments
AuthorsOpenReview ↗

Contribution over prior work

We thank the reviewer for the constructive comments and address each point below.

W1. The framework builds on existing spherical flow matching, visual retrieval, and attention mechanisms, with the main additions concentrated in contextual conditioning and the geodesic distance bias. The paper should distinguish these contributions more clearly from prior retrieval-augmented geolocation methods and explain what modeling capabilities they add beyond standard conditioning.

Q1. What are the main modeling advances over existing spherical flow matching and retrieval-augmented geolocation methods? What does the geodesic distance bias add beyond ordinary conditional attention?

Over spherical flow matching. Dufour et al. (2025) model pθ(yx∣x)p _ \theta(\mathbf{y} _ x \mid \mathbf{x}), so all geographic knowledge is stored in the parameters of the flow matching model. We model pθ(yx∣x,Cx)p _ \theta(\mathbf{y} _ x \mid \mathbf{x}, \mathcal{C} _ x), where the context Cx\mathcal{C} _ x is retrieved from an external database; we enable adding extra knowledge during test time. The benefit of our context module is clear; the velocity network, descriptors and training recipe remain the same in our experiments, meaning it is a clean and fair comparison.

Over retrieval-augmented geolocation. The closest retrieval-augmented predictors to ours are Img2Loc, G3 and GeoToken. GeoRanker, in contrast, is primarily a retrieval-based method; its contribution is better re-ranking of candidates, not coordinate prediction. Img2Loc and G3 rely on a frozen MLLM and its general world knowledge, and they produce coordinates as generated text. We instead train a dedicated module that predicts directly in continuous coordinates. GeoToken works in a quantized space of globe cells and predicts discrete cell labels. We operate on continuous coordinates instead, which lets us encode the geodesic distances to the context images directly in the bias term. Table 3 shows that this design is effective.

Geodesic distance bias. The attention weights without the bias cannot see where the current sample yty_t lies relative to each context image. The bias supplies this geodesic distance directly, with a learned, time-dependent coefficient. Encoding locations in token space instead did not improve over the model without the bias. Removing the bias worsens every metric, e.g. city accuracy from 26.1 to 23.3% (Table 3 of the paper).

We will make these distinctions explicit in the related work of the final paper.

AuthorsOpenReview ↗

Retrieval design

W2. Retrieval relies on frozen pretrained descriptors and cosine similarity, with little task-specific design. There is room to improve the retrieval representation or matching mechanism through local feature correspondences, geometric consistency, and greater robustness to viewpoint and lighting changes, reducing reliance on global appearance similarity.

Q2. What motivates the choice of frozen descriptors and cosine retrieval? How could local geometric information and invariance to viewpoint and lighting changes be incorporated into a more task-specific retrieval mechanism?

Frozen pretrained descriptors. Our contribution is in how the retrieved images are used, so we deliberately kept the retrieval itself as simple as possible. We agree that there can be performance improvements by using better descriptors that result in better retrieval. The results in Table 4 of Appendix in the original submission already demonstrated this. Cosine similarity over these descriptors is the standard, efficient choice, and exact search over the database takes only 0.2 ms per query as shown in Table B3.

Using local descriptors for re-ranking in retrieval. Thank you for the excellent suggestion. We performed a new experiment. We re-rank the K=64K=64 context images of each query with a learned similarity over local descriptors from DINOv3 using the recent re-ranker ELViS [1]. The models of the paper (StreetCLIP on OSV-5M, PE-Core-G on the Flickr benchmarks) are evaluated on the new image order within the same context without retraining, on the full test sets. Tables B1 and B2 show that re-ranking improves RAFMa on all settings (there is a tiny exception). It also improves the 1-NN baseline, but RAFMa is still better. This experiment will be added to the final paper.

[1] ELViS: Efficient Visual Similarity from Local Descriptors that Generalizes Across Domains, Suma et al., ICLR 2026

Table B1. Re-ranking the context with local descriptors on OSV-5M, without retraining, with StreetCLIP descriptors and 210,122 queries. The K=64K=64 context images of each query are retrieved as in the paper and re-ordered by a learned similarity over local DINOv3 descriptors. 1-NN predicts the location of the first context image.

MethodGeoScore ↑Dist. (km) ↓Country (%) ↑Region (%) ↑Subreg (%) ↑City (%) ↑
1-NN3645133774.647.129.220.8
1-NN, local re-ranking3873108278.853.835.125.4
RAFMa (paper)406285882.858.437.626.1
RAFMa, local re-ranking409982883.559.939.527.7

Table B2. Re-ranking the context with local descriptors on the Flickr benchmarks, as in Table B1, with PE-Core-G descriptors: YFCC4k (4,536 queries) and Im2GPS3k (2,997 queries). Accuracy (%) within the indicated distance in km (↑).

MethodYFCC4k 1 kmYFCC4k 25 kmYFCC4k 200 kmYFCC4k 750 kmYFCC4k 2500 kmIm2GPS3k 1 kmIm2GPS3k 25 kmIm2GPS3k 200 kmIm2GPS3k 750 kmIm2GPS3k 2500 km
1-NN34.545.755.367.680.117.844.158.573.185.8
1-NN, local re-ranking38.848.356.166.679.120.846.758.771.784.3
RAFMa (paper)39.856.168.480.489.118.852.070.585.193.6
RAFMa, local re-ranking41.956.868.780.689.220.752.870.884.893.6
AuthorsOpenReview ↗

Cost of the full pipeline

W3. The number of trainable parameters does not establish the efficiency of the full system. The reported 135M parameters exclude the frozen visual encoder, while inference also requires retrieval from a large feature database and 250 guided sampling steps. Without comparisons of end-to-end inference time, peak GPU memory, and database storage, the practical accuracy–cost trade-off remains unclear.

Q3. Please report the full pipeline's inference time, peak GPU memory, and database storage requirements, including visual encoding, retrieval, and generation, and discuss how accuracy and cost change with database size and sampling steps.

Runtime and memory. Table B3 reports the time and GPU memory of each stage of the pipeline, with PE-Core-G descriptors and the MP-16 database on a single A100 GPU. For RAFMa, the frozen image encoder accounts for 96% of the time, a cost shared by every method built on the same encoder. GeoRanker is about 37 times slower than RAFMa. Our pipeline peaks at 22.9 GB in the retrieval stage, which holds the database on the GPU. The weights of RAFMa take 0.5 GB, compared to 16.6 GB for the GeoRanker re-ranker. The table will be added to the final paper.

Accuracy and runtime vs steps. The 250 steps were inherited from Dufour et al. (2025) and are more than needed. In Figure 13 (Appendix C), accuracy saturates within 32 steps, while generation time grows linearly with the number of steps. In Table B3 we therefore report the generation cost of RFM and RAFMa at 32 steps, where accuracy is on par with 250 steps on every metric.

Database size. Storage and retrieval time grow linearly with the number of database images, and accuracy keeps rising with it (Figure 4 of the original submission). However, the database descriptors can also be compressed. Table B4 shows that, stored as their first 512 principal components with 4 bits each, the descriptors take 1.2 GB instead of 12.1 GB on MP-16, with insignificant accuracy loss. Retrieval uses the compressed descriptors and the model reads their reconstructions in the original dimensionality, PE-Core-G on all three benchmarks.

Table B3. Cost of each method on one GPU (NVIDIA A100, 40 GB), with PE-Core-G descriptors and the MP-16 database. Time is in ms per image, in batches of 64 for the encoder, 1024 for retrieval and generation and 8 queries for GeoRanker. RFM and RAFMa use 32 Euler steps. GPU memory (GB) is the peak of each stage. The total memory is the largest of the stage peaks. Image decoding is not included, nor is the CPU preprocessing of GeoRanker (268 ms per image).

Stage1-NN time (ms)1-NN memory (GB)RFM time (ms)RFM memory (GB)RAFMa time (ms)RAFMa memory (GB)GeoRanker time (ms)GeoRanker memory (GB)
Image encoder29.76.729.76.729.76.729.76.7
Retrieval0.2222.9––0.2222.90.2222.9
Generation / re-ranking––0.930.61.184.21,10732.0
Total29.922.930.66.731.122.91,13732.0

Table B4. Compressing the database descriptors with PCA. The PE-Core-G descriptors of the database images are stored as their first dd principal components, with 4 bits per component. Retrieval and the model both use only the compressed descriptors. Storage is computed for the 4.72 M images of MP-16.

ddStorage (GB)YFCC4k GeoScoreYFCC4k 25 km (%)Im2GPS3k GeoScoreIm2GPS3k 25 km (%)OSV-5M GeoScoreOSV-5M 25 km (%)
640.15389648.3416049.0425925.3
1280.30397652.7418650.2429430.4
2560.61402254.8419151.6432533.6
5121.21403055.8421452.0433734.7
1280 (full)12.09403356.1421352.0434235.0

Thank you for the clarifications and additional experiments. My concerns have largely been addressed. Please include the full runtime, memory, and database storage results in the paper, and discuss the additional memory required by retrieval. I will maintain my original score.

Reviewer_XogdOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RAFMa augments spherical flow matching for global visual geolocation with retrieved image–location pairs. A context network processes visual descriptors and aggregates geographic information through cross-attention. Its attention logits incorporate a learned, time-dependent geodesic distance bias between the current generated location and retrieved locations, allowing context weighting to evolve during sampling. The image encoder remains frozen, while the context and velocity networks are trained jointly. The paper reports improvements over a context-free RFM baseline across multiple encoders, demonstrates benefits from expanding the retrieval database without retraining, and presents competitive benchmark results. Final predictions may additionally be snapped to nearby retrieved coordinates.

Critical strengths
  1. Well-motivated integration of retrieval and generation. Conditioning continuous geographic generation on external image–location pairs addresses a useful combination of geographic generalization and database-based knowledge updates.
  2. Meaningful empirical improvements with a controlled backbone. RAFMa improves over the context-free baseline across all tested encoders on OSV-5M. Matching the velocity-network architecture and training recipe strengthens the evidence for the value of contextual conditioning.
  3. Useful evidence of inference-time adaptability. Database expansion and improved retrieval benefit a fixed trained model. The reduced-training-data experiments also demonstrate the utility of external memory, while retaining access to the full labeled database.
Critical weaknesses
  1. Unquantified training–test retrieval gap. Training excludes the query itself, whereas OSV-5M additionally separates test queries from the database by 1 km. Without training-set retrieval statistics, it is unclear whether the model learns from substantially easier contexts than those encountered at test time.
  2. Missing analysis for the default context size. The default configuration uses 64 context images, but Figure 5 and Table 6 analyze only 32. The paper does not explain this discrepancy or quantify retrieval coverage and sensitivity to distractors under the main experimental setting.
  3. Insufficient isolation of the FM-based fusion contribution. The ablations establish improvements within an RFM architecture but do not compare against simpler predictors using the same context. Fine-scale results also depend substantially on coordinate snapping, leaving the benefit of integrating retrieval throughout generation insufficiently separated from post-processing.
Questions for authors
  1. Could the authors report training and test retrieval recall at multiple geographic thresholds under (K=64), together with nearest-context distance distributions and any training-time geographic or sequence exclusions?
  2. Why do the context analyses use (K=32)? Could these analyses be extended to the default (K=64), including a controlled comparison of retrieval recall and localization accuracy?
  3. With retrieval and post-processing fixed, how does RAFMa compare with a simpler context-conditioned predictor and with context-free RFM followed by identical coordinate snapping? Please use comparable parameter budgets where applicable.
5 comments
AuthorsOpenReview ↗

Training–test retrieval gap

We thank the reviewer for the careful review and the concrete suggestions, and address each point below.

W1. Unquantified training–test retrieval gap. Training excludes the query itself, whereas OSV-5M additionally separates test queries from the database by 1 km. Without training-set retrieval statistics, it is unclear whether the model learns from substantially easier contexts than those encountered at test time.

Q1. Could the authors report training and test retrieval recall at multiple geographic thresholds under (K=64K=64), together with nearest-context distance distributions and any training-time geographic or sequence exclusions?

The reviewer is right, there is a discrepancy between training and testing on OSV-5M.

To keep the same training setting across datasets, in the original submission, we had not excluded images within 1 km from the query in the training of OSV-5M. However, we perform this experiment and report the retrieval statistics in Table A1. Indeed, the mentioned discrepancy results in differences between these statistics (row 1 vs row 3). On MP-16 there is no such discrepancy (rows 4–6).

Excluding images within 1 km of the query in training reduces the difference to the test statistics (row 2 vs row 3). Nevertheless, a model trained with this exclusion (i.e. trying to mimic the test distribution in the training set), performs similarly to the original training, as presented in Table A2. This is good news. Our model and training are not sensitive to the discrepancy.

We will integrate the analysis along with the additional results to the appendix of the paper.

Table A1. Retrieval statistics. Each row gives the share of queries (%) for which at least one of the 64 context images lies within the given distance of the true location. The median distance is reported to the top-ranked context image and to the nearest of the 64 context images. Rows in bold are the training contexts of the models in the paper.

RowDataset, descriptorsQueriesExcluded: queryExcluded: within 1 km≤ 1 km≤ 25 km≤ 200 km≤ 750 km≤ 2500 kmMedian top (km)Median nearest (km)
1OSV-5M, StreetCLIPTraining✓80.797.499.599.9100.01.450.20
2OSV-5M, StreetCLIPTraining✓✓0.093.998.999.799.912.21.49
3OSV-5M, StreetCLIPTest✓✓0.055.684.895.298.715218.5
4MP-16, PE-Core-GTraining✓54.274.490.197.599.51970.55
5MP-16, PE-Core-GYFCC4k58.874.889.096.099.080.70.20
6MP-16, PE-Core-GIm2GPS3k48.075.492.397.699.572.61.21

Table A2. Excluding images within 1 km from the query on OSV-5M training, with StreetCLIP descriptors. Both models use the same training recipe and are evaluated as in Table 1 of the paper. Only the context of the training queries changes.

Training contextGeoScore ↑Dist. (km) ↓Country (%) ↑Region (%) ↑Subreg (%) ↑City (%) ↑
Only query excluded (paper)406285882.858.437.626.1
Everything within 1 km excluded406886682.858.737.825.7
AuthorsOpenReview ↗

Analysis at the default context size

W2. Missing analysis for the default context size. The default configuration uses 64 context images, but Figure 5 and Table 6 analyze only 32. The paper does not explain this discrepancy or quantify retrieval coverage and sensitivity to distractors under the main experimental setting.

Q2. Why do the context analyses use (K=32K=32)? Could these analyses be extended to the default (K=64K=64), including a controlled comparison of retrieval recall and localization accuracy?

Figure 5 and Table 6 in the original paper were produced with exactly the same model as all other results, and the analysis reaching only K=32K=32 was an oversight. We have updated the original manuscript with the corrected versions using K=64K=64 for both, and also for the qualitative examples in Figures 3, 7, and 8. The observations and conclusions are unchanged. Please note that the example (c) of Figure 8 was replaced by a different query, as the previous one no longer has zero context images near the true location with K=64K=64.

To verify the effect of the context size KK, we vary the number of context images at inference and show the accuracy next to the context quality in newly added Figure 10 in Appendix B.5. Accuracy and context quality both increase with KK up to the context size used in training. Beyond it, the context quality keeps increasing while the accuracy saturates, and degrades on OSV-5M. The model trained using K=64K=64 shows slight improvement over the one trained using K=32K=32 in most settings.

Finally, Table A3 quantifies the sensitivity to distractors with K=64K=64. The accuracy is correlated with the number of context images near the true location, while larger distractor groups make performance worse and improvement compared to the baseline smaller. Nevertheless, the improvements compared to the baseline are overall significant whenever at least one context image is within 25 km. The table will be added to the final paper.

Table A3. Sensitivity to distractors with K=64K=64. Test queries are divided into different categories by the number of context images within 25 km of the true location (rows) and by the size of the distractor group (columns). The distractor group is the largest set of context images that are farther than 25 km from the true location and within 25 km of each other. Table cells report the accuracy at 25 km and, in parentheses, the percentage of test queries of each category, in %. RAFMa and the context-free baseline on OSV-5M with StreetCLIP descriptors, without post-processing.

Context images within 25 kmRAFMa, 0–8RAFMa, 9–16RAFMa, ≥ 17Baseline, 0–8Baseline, 9–16Baseline, ≥ 17
01.3 (10.1)1.0 (18.1)0.8 (16.2)3.2 (10.1)3.8 (18.1)4.1 (16.2)
128.2 (2.1)26.0 (3.8)21.9 (3.0)10.5 (2.1)9.7 (3.8)9.1 (3.0)
2–451.6 (2.9)45.2 (5.5)37.7 (4.6)23.3 (2.9)19.0 (5.5)17.5 (4.6)
5–1677.8 (3.8)70.8 (8.3)60.5 (7.6)44.2 (3.8)40.5 (8.3)36.4 (7.6)
≥ 1796.1 (4.5)90.4 (5.5)82.2 (4.0)78.1 (4.5)70.1 (5.5)66.4 (4.0)
AuthorsOpenReview ↗

Isolating the flow-matching contribution (1/2)

W3. Insufficient isolation of the FM-based fusion contribution. The ablations establish improvements within an RFM architecture but do not compare against simpler predictors using the same context. Fine-scale results also depend substantially on coordinate snapping, leaving the benefit of integrating retrieval throughout generation insufficiently separated from post-processing.

Q3. With retrieval and post-processing fixed, how does RAFMa compare with a simpler context-conditioned predictor and with context-free RFM followed by identical coordinate snapping? Please use comparable parameter budgets where applicable.

Following the reviewer's suggestions, we trained a simpler predictor that uses the same context as RAFMa, and a context-free RFM with a comparable number of parameters. Their results are in Tables A4 and A5. Table A6 compares RAFMa and the context-free RFM with identical post-processing.

Simpler predictor with the same context conditioning as RAFMa. We train a regression model that keeps the context network and the AdaLN-MLP of RAFMa, but outputs the location directly in a single forward pass. The training hyperparameters, retrieval part, and post-processing are fixed for all setups, with StreetCLIP descriptors on OSV-5M and PE-Core-G on the Flickr benchmarks. It uses the same training recipe as RAFMa and is trained to minimise the geodesic distance to the true location, the best of three regression losses we tried. Tables A4 and A5 report the results on OSV-5M and on the Flickr benchmarks, respectively. RAFMa outperforms the simple regression predictor on every metric besides the coarsest setting on the Flickr benchmark. Context improves the regression baseline on OSV-5M and YFCC4k as well, but less than in RAFMa.

Baseline RFM with larger model to match the number of parameters of RAFMa. In Tables A4 and A5, we observe that increasing the model size (by the number of the additional parameters of the context module) does not bring any significant improvement. Therefore, the improvement of RAFMa comes from the context and its use during generation, not from the additional parameters.

Table A4. RAFMa against simpler predictors on OSV-5M, with StreetCLIP descriptors. Every model receives the same K=64K=64 context images and the same post-processing. RFM S2\mathcal{S} _ 2 is the context-free baseline, and matched size is the same model widened to the size of RAFMa. Regression outputs the location directly, in a single forward pass.

ModelParams (M)GeoScore ↑Dist. (km) ↓Country (%) ↑Region (%) ↑Subreg (%) ↑City (%) ↑
Regression, context-free134.1392092678.948.224.212.6
Regression, with context133.1391794978.650.128.718.0
RFM S2\mathcal{S} _ 2119.6398593581.054.231.117.5
RFM S2\mathcal{S} _ 2, matched size135.0400091881.154.431.217.6
RAFMa (paper)134.6406285882.858.437.626.1

Table A5. RAFMa against simpler predictors on the Flickr benchmarks, as in Table A4, with PE-Core-G descriptors. Accuracy (%) within the indicated distance in km (↑).

ModelParams (M)YFCC4k 1 kmYFCC4k 25 kmYFCC4k 200 kmYFCC4k 750 kmYFCC4k 2500 kmIm2GPS3k 1 kmIm2GPS3k 25 kmIm2GPS3k 200 kmIm2GPS3k 750 kmIm2GPS3k 2500 km
Regression, context-free134.317.628.458.078.088.29.533.066.786.193.9
Regression, with context133.320.134.160.479.389.48.730.465.984.793.4
RFM S2\mathcal{S} _ 2119.819.837.659.777.688.416.548.770.284.993.6
RFM S2\mathcal{S} _ 2, matched size135.219.437.759.977.688.415.948.569.684.993.2
RAFMa (paper)135.039.856.168.480.489.118.852.070.585.193.6

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank all reviewers for their valuable feedback. We answer the questions and address the weaknesses in comments under each review. The manuscript's appendix now includes the figures we refer to within our responses. Changed or added text is highlighted in red, and for a redrawn figure its caption is highlighted. Additional results are reported as tables within our responses. The manuscript was mainly updated to support the responses and should not be seen as a new final version. We are happy to provide any further analysis or clarification during the discussion. Afterwards, we will update the manuscript with the new results; within each response we state what will be added.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.